Jul 25, 2026 · 1h 2m · 20vc
Mercor Head of Product on Revenue Concentration from Frontier Labs
Clips from this episode (6)
Short vertical cuts produced from the tape, captions burned in. Where a cut lands on a statement from the ledger, its card says so.
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of 20VC, Mercor Head of Product Osvald Nitski discusses the economics of frontier AI data, the evolving role of product managers in AI-native companies, and strategies for scaling enterprise AI workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 26.6% of the talking time here. How this is scored →
Harry as informed peer 5.8 Guest teaching 4.9 Guest disagreement 3.1 Harry pushing back 4.6
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 2:30 Osvald rejects 90/10 workflow consensus
Osvald directly dismisses conventional calculations that 90% of enterprise workflows will be automated by open models, arguing they ignore vast latent demand.
Hardest push from Harry ▶ 1:00:50 Harry attacks teleoperated robotics hypeHarry aggressively rejects Osvald's excitement about physical robotics, mocking demos where teleoperators secretly control sluggish household machines.
Biggest teaching moment ▶ 44:24 Osvald explains VC-subsidized founder annotationOsvald breaks down the unscalable unit economics of founder-led annotation boutiques offering below-market labor subsidized by VC funding.
Harry holds his own ▶ 25:11 Harry demands concrete substance behind hiring buzzwordsHarry forcefully interrupts Osvald's abstract phrasing around systems design and statistical experimentation to demand practical interview questions.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Open Source Models vs. Core Data Services | 6 | 6 | 5 | 6 | Harry opens by probing whether open source models cannibalize Mercor's core business and challenges the addressable market for the remaining 10% of workflows. Osvald pushes back directly on the premise, rejecting the 90/10 framing and citing latent demand for long-horizon agentic tasks. | |
| Enterprise Data Privacy and Model Sensitivity | 5 | 6 | 4 | 5 | Harry points out the apparent irony of enterprises running sensitive data on open-source weights while using closed models for HR. Osvald educates him on how open weights allow local inference control and explains why continuous uncapped reward tasks cannot be measured with binary completion percentages. | |
| Specialized Enterprise Models and Data Demand | 6 | 5 | 3 | 4 | Harry brings up perspectives from Fireworks and Palantir's Alex Karp to question enterprise AI ROI. Osvald calmly reframes the current market state from an ROI crisis to an ongoing exploratory phase where companies accept experimental spend. | |
| Developer Token Budgets vs. Salary Ratios | 6 | 4 | 2 | 4 | Harry cites specific figures from Salesforce and Mercor founder Brandon regarding developer token budgets versus salaries. Osvald agrees with the trend, clarifying that hypergrowth dynamics make heavy token spend economical. | |
| The Evolving Role of Product Managers in AI | 5 | 5 | 3 | 3 | Harry asks whether companies simply build 10x more products or reduce team sizes in the AI era. Osvald explains that product management's core challenge becomes ruthlessly limiting surface area while shifting from tool proficiency to pure business judgment. | |
| Product Mistakes and Setting Operational Guardrails | 5 | 4 | 2 | 5 | Harry presses Osvald to reveal concrete past mistakes and how Mercor determines enduring demand. Osvald candidly recounts over-engineering their annotation tool for every customer request before instituting strict operational guardrails. | |
| Consolidating Toolchains and Moving Away from Figma | 7 | 4 | 3 | 6 | Harry brings up quotes from Factory's Matan arguing that services are an excuse for crap software, challenging the role of forward deployed engineers. Osvald counters that services bridge the current knowledge dissemination gap while talent matures. | |
| Modern AI Hiring Strategies and Whiteboard Testing | 6 | 4 | 3 | 6 | Harry challenges Osvald on whether Mercor is succumbing to VC-led hiring tropes by seeking seasoned operators. He also cuts in to demand concrete definitions of vague interview buzzwords like systems design and statistical experimentation. | |
| Preserving Human Creativity and Decision Muscle | 5 | 4 | 2 | 3 | Harry draws parallels to venture investing where extensive diligence does not always correlate with positive outcomes. Osvald emphasizes drawing a strict boundary between using AI for execution and retaining human ownership over decision judgment. | |
| Product Pod Structures and Operational Cadence | 5 | 5 | 2 | 3 | Harry inquires about team topology, PM-to-engineer ratios, and meeting cadences. Osvald breaks down Mercor's pod structure across Studio and Marketplace, explaining why engineering velocity will increase PM headcount ratios. | |
| Marketplace Supply Scaling and Unit Economics | 6 | 5 | 3 | 4 | Harry asks if high payout rates are the sole secret to supply retention and queries gross margins. Osvald clarifies that sustainable retention requires consistent work flow and dignified expert experience rather than short-term payout spikes. | |
| Addressing Revenue Concentration and Down-Market Motion | 6 | 5 | 3 | 5 | Harry presses on customer concentration risks given Mercor's heavy reliance on top frontier labs. Osvald details the product roadmap to move down-market and automate project management so non-lab enterprises can self-serve data creation. | |
| Operational Intensity and Reinforcement Learning Environments | 5 | 6 | 3 | 3 | Harry asks what makes democratizing data annotation so hard and queries upcoming demand trends. Osvald explains the massive operational paranoia required for edge-case alignment and breaks down why reinforcement learning environments are the new frontier. | |
| Lab Price Sensitivity vs. Founder-Led Annotation | 6 | 6 | 4 | 4 | Harry asks if labs aggressively negotiate on pricing. Osvald outlines the lab procurement dynamic and highlights how Mercor competes with a cottage industry of VC-subsidized, founder-led annotation shops that fail to scale. | |
| Competitive Intelligence and Market Focus | 6 | 5 | 3 | 5 | Harry discusses portfolio companies like Leya competing against Anthropic and questions big tech distraction. Osvald draws historical parallels to Google and Microsoft launching non-core initiatives that ultimately lost to focused category specialists. | |
| Cybersecurity Data and Adversarial AI Training | 5 | 6 | 3 | 4 | Harry brings up Mercor's previous security incident and asks about rising vulnerability risks from AI code generation. Osvald explains that cybersecurity represents a unique adversarial data game where completion targets continually move. | |
| San Francisco Talent Competition and Culture | 6 | 3 | 3 | 5 | Harry challenges the San Francisco talent market and asks whether high-performing difficult personalities are tolerated. Osvald affirms that extreme talent and ownership outweigh abrasive interpersonal quirks. | |
| Quickfire Round: Career Advice, Robotics, and Vision | 8 | 5 | 4 | 8 | In the quickfire round, Harry strongly pushes back on Osvald's bullishness on robotics, labeling home robot demos clunky and questioning the scalability of Waymo compared to global transport needs. Osvald defends the comparison by citing Waymo's ten-year developmental arc. |
Statements from this episode (35)
Nitski: Material learned in school will quickly become outdated
“Get a real internship as soon as possible because whatever you learn in school is probably going to be outdated quickly.”
Nitski: Open-source AI does not cannibalize frontier data demand
“I wouldn't say that open source model improvements cannibalize our core business because data is most valuable on the frontier of model performance, so each of our customers has their own unique goals and is purchasing eval and training data sets To fill gaps …”
Nitski: Top AI models cannot handle 90% of enterprise workflows
“I'm not convinced that 90% of enterprise workflows can be handled by open models or Frontier models right now.”
Nitski: Top AI models score 50% on Mercor's long-horizon benchmarks
“Well, in our Apex benchmarks, we're getting closer to around 50% of Long Horizon workflows. Top models are scoring around around that much.”
Nitski: Specialized AI models will require custom enterprise training data
“I buy it. I think it's also self-serving towards Mercore, in that we think that every specialized model will need enterprise specific eval and training data to show the model how it's performing in its setting, and I think it depends on I think the diversity a…”
Nitski: Enterprise AI does not currently have an ROI problem
“I don't think there's an ROI problem right now. I think we're in a period of exploration and experimentation where there's more tolerance, more patience To get that ROI calculation right now.”
Nitski: judge token spend by outcomes, not one salary ratio
“I hope that we can move towards a future of better accounting of the outcomes being driven by token spend. Because even here, I think in a company like Salesforce we have so many, a company of that size, certainly you're getting, you should have different spen…”
Nitski: Enterprise AI token spend will exceed 3% of developer salaries
“I think macro, the percentage will increase over time to more than three percent.”
Nitski: Mercor cannot spend money fast enough to meet customer demand
“We can't spend money fast enough to service all of the demand that we have.”
Nitski: AI is driving higher PM-to-engineer headcount ratios
“So the trend that we see is we're, as a product team, constantly fighting to reduce surface area and simplify things, and we also see a higher ratio of PMs to ENG because engineering is less bottlenecked.”
Nitski: AI removes skill barriers, making judgment the key PM differentiator
“I can do things very quickly now, you know, it's like there's, ah, skill issues have almost gone away. So now it's all about, ah, judgment, and am I doing what is going to drive the most business value?”
Nitski: Mercor should have restricted supported annotation workflows much sooner
“We made a tool that's maximally flexible, has all sorts of, we had, like, hundreds of different projects running on it. That's just chaos to manage, and what we needed to do sooner was to put guardrails on the type of services that we support, and work closer …”
Nitski: Mercor team is shifting away from Figma to Claude
“To be honest, I let the team do whatever is best for them, and this is a trend I've just observed amongst almost everybody, is that cloud design has done a great job. People really like using it, it's easy to use, and we've just had a natural movement towards …”
Nitski: Forward-deployed AI services are only a short-term enterprise trend
“I think it's the future for the short term as the knowledge of how to use AI gets disseminated throughout industry. We have basically a concentration of a bunch of people in San Francisco who really know how to deploy agents, eval agents, Be AI first in engine…”
Nitski: Enterprise AI agent deployment talent will take a decade
“I think that it's a knowledge dissemination problem, so I think the, that, that's one way to look at it. The other way is, why not hire someone to just do this agent deployment at your own company? And I just don't think the skill is out there yet. I don't thi…”
Nitski: All core team attrition at Mercor is to found companies
“I'm proud that of the people I work closest with on my teams, I've only had attrition to founding. And we've had quite a bit of it.”
Nitski: Mercor shifts to senior hires as AI obsoletes junior skills
“We've biased towards more senior hires. Who are better at understanding what drives the business forward, finding, kind of, like, really crocking how we operate, how we make more revenue, how we deliver better services to our customers, how we keep our custome…”
Mercor replaces take-home coding assignments with AI agent fluency tests
“We've moved away from take-home assignments. We do one take-home assignment, which is, like, can you just, like, use an agent to go, you're on your own for a bit of time, go use an agent, you know, give, produce this artifact for me, and we'll look at it.”
Stebbings: I write all 20VC questions myself and refuse AI
“I do all questions myself. I would never use AI and I'm very concerned about it because you lose the muscle to me.”
Nitski: Teams must never delegate core decision-making to AI
“I want very careful never to delegate judgment or decision making to models because it's, they make you think that it's doing the right thing, but You have to be paranoid with them still, right? You still have to, like, double check everything, and that's what…”
Nitski: Mercor's total headcount grew over 10x in the last year
“As the headcounts increased, you know, like more than 10 X in the last year, we just want to be careful not to grow one faster than the other.”
Nitski: Mercor adds millions of dollars in cash to bank weekly
“It doesn't annoy me, no, because the, you know, we end every week with millions more in the bank, right? So it's funny how you can have, I've been at other Companies where I've seen, you know, interesting financial engineering and accounting, and people can ha…”
Nitski: Mercor's fix for lab concentration is moving down market
“I can answer this from kind of like a, how it affects the product team. We would love to move like our biggest challenge is moving down market so that every single enterprise can efficiently run human data projects for eval and training. And that'll diversify …”
Nitski: Every enterprise will eventually require human data work
“It's the direction we have been heading, which has reduced concentration, and it's the direction that we'll continue to head as every enterprise begins to have human data work for their proprietary use cases.”
Mercor shifts human data services to RL environments across modalities
“The data types also change very frequently. So we're, we've moved from supervised fine tuning to preference ranking to Rubric based annotation to now RL environments across a whole bunch of different modalities.”
Nitski: RL environments are Mercor's fastest-growing data modality
“The data type that's growing the fastest for us is environments.”
Nitski: niche data providers exist already, and the small ones struggle
“To an extent, we're already in this world. It's not that successful though for the small players always. So how I would describe it is we're facing what looks like a cottage industry of founders doing annotation themselves, right?”
Nitski: Founder-led AI data startups cannot scale to enterprise lab demands
“This is just, like, VC-subsidized work that labs love. The problem is scaling it beyond a few data points or what one founder or full-time employees can do, and this is the position that we're in, is we're having to compete against Basically founder-led annota…”
Nitski: Mercor sees rapidly growing demand for cybersecurity AI data
“We see very Rapidly increasing demand for cyber defensive capabilities via data, and very interesting data types”
Nitski: AI will never automate 90% of cybersecurity workflows
“There's never going to be that 90% for security because the goal posts are always going to move.”
Nitski: Evals and training data are AI's primary bottleneck
“Evals and training data are the primary bottleneck to model performance right now.”
Nitski: Physical and robotics data for AI will grow significantly
“I think that real world, like physical data is going to grow significantly over the next three years.”
Nitski: Robotics is the most underhyped tech opportunity three years out
“Probably the same answer as before in that, like, the three years out opportunity of robotics.”
Stebbings: Autonomous vehicles remain largely irrelevant and limited
“I love it, but it's in one city. It can't deal with, like, very ambiguous data. It's still, like, pretty irrelevant.”
Nitski: Robotics adoption will scale like driverless cars, not ChatGPT
“Yeah, I think so, but I think it might play out similar to driverless cars, where it's really hard to scale physical things as opposed to software, so it might be more of a, more of like a Waymo robo-taxi-cruise type moment than a ChatGPT moment, but I think t…”